Build Simpler Production ML Systems using Feature/Training/Inference Pipelines
summary
Claims a unified Feature Training Inference pipeline pattern and shows a working loan‑approval demo built on an open feature store and model registry: feature groups with event time for point‑in‑time joins, versioned feature groups and feature views to keep train/inference consistent, online vs offline stores, and KServe for serving. Code uses pandas notebooks, inserts data into the online feature store, trains models, and serves a real‑time UI; online retrieval benchmarks are single‑digit ms and the example model initially over‑approves loans.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 5:25Why ML systems are hard
- 11:41Loan approval project setup
- 19:32Feature pipeline
- 33:20Training with feature views
- 46:16Batch and real-time inference
- 55:42Versioning and breaking changes
- 1:00:17Q&A
related talks
36mBuild prod ML systems with feature /training /inference pipelines
01h 52mDeploying a Fraud Detection Model with a Feature Store
39mMachine Learning in Financial Credit Risk Assessment
01h 28mFeature Engineering Made Simple
38mReal Time ML Pipelines in Multi-Tenant Environments
29mMachine learning in production
30mHow Feature Stores Enable Operational ML
17mCreating & operating ML models from event-based data using feature stor...